import streamlit as st from transformers import CLIPProcessor, CLIPModel import torch from PIL import Image import requests from io import BytesIO !pip install --upgrade torch st.title("Text to Image Generation with CLIP") # Load pretrained models clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch16") clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch16") text = st.text_area("Enter a description:") if st.button("Generate Image") and text: # Process text and get CLIP features for text text_features = clip_processor(text, return_tensors="pt", padding=True) # Load an example image from the web (replace this with your image loading logic) example_image_url = "https://source.unsplash.com/random" example_image_response = requests.get(example_image_url) example_image = Image.open(BytesIO(example_image_response.content)) # Process image and get CLIP features for image image_features = clip_processor(images=example_image, return_tensors="pt", padding=True) # Ensure the dimensions of pixel_values are the same for text and image features max_len = max(text_features['pixel_values'].shape[1], image_features['pixel_values'].shape[1]) text_features['pixel_values'] = torch.nn.functional.pad(text_features['pixel_values'], (0, max_len - text_features['pixel_values'].shape[1])) image_features['pixel_values'] = torch.nn.functional.pad(image_features['pixel_values'], (0, max_len - image_features['pixel_values'].shape[1])) # Concatenate text and image features combined_features = { "pixel_values": torch.cat([text_features['pixel_values'], image_features['pixel_values']], dim=1) } # Forward pass through CLIP image_representation = clip_model(**combined_features).last_hidden_state.mean(dim=1) # For visualization, you can convert the image representation back to an image image_array = image_representation.squeeze().cpu().numpy() generated_image = Image.fromarray((image_array * 255).astype('uint8')) # Display the generated image st.image(generated_image, caption="Generated Image", use_column_width=True)